How AI is trained, and why different models feel different

Ever wondered why ChatGPT feels different from Claude, or why some AI tools seem friendlier while others feel more clinical? It comes down to how they were trained. In this second video of the AI for Business series, 25eight CEO Sam Hurley walks through how large language models are built, from the billions of data points they learn from, to the human reviewers who shape their behaviour, to the guardrails organisations add before you ever type a prompt.

The video covers the two levels of training that shape every response: global training done by the organisation behind the model, and personal shaping done by you through context and instructions. The global side runs through four stages: pre-training on huge volumes of text, alignment through human feedback, guardrails and safety systems, and personalisation. Each company makes different choices at each stage, which is why the tools have distinct personalities and strengths. The practical takeaway is that AI output is a statistical prediction shaped by training choices rather than a neutral source of truth. Once that sinks in, you start asking better questions of it.

Key points

• LLMs predict language from patterns, in a way loosely similar to how your brain completes a familiar phrase.
• Models are shaped in four stages: pre-training, alignment, guardrails and personalisation.
• Human reviewers and company choices give each model its distinct personality.
• No single tool is best for everything; the right one depends on the task.
• Treat AI output as prediction to be checked rather than settled fact.

This video expands on the training section of An introduction to AI for small business.